vect-patterns: Fix up vect_recog_rotate_pattern [PR108064]

Hi!

Since vect_recog_rotate_pattern has been extended to work also
on signed types in r13-1100 we miscompile the testcase below.
vect_recog_rotate_pattern actually emits correct scalar code into
the pattern def sequence (in particular cast to utype, doing the
2 shifts in utype so that the right shift is logical and not arithmetic,
or and then cast back to the signed type), but it didn't supply vectype
for most of those pattern statements, which means that the generic handling
fills it up later with the vectype provided by vect_recog_rotate_pattern.
The problem is that it is vectype of the result of the whole pattern,
i.e. vector of signed values in this case, while the conversion to utype,
2 shifts and or (everything with utype lhs in scalar code) should have
uvectype as STMT_VINFO_VECTYPE.

Fixed with following patch, bootstrapped/regtested on x86_64-linux and
i686-linux, ok for trunk?

2022-12-13  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/108064
	* tree-vect-patterns.cc (vect_recog_rotate_pattern): Pass uvectype
	as 4th argument to append_pattern_def_seq for statements with lhs
	with utype type.

	* gcc.c-torture/execute/pr108064.c: New test.


	Jakub

> Am 13.12.2022 um 10:28 schrieb Jakub Jelinek via Gcc-patches <gcc-patches@gcc.gnu.org>:
> 
> Hi!
> 
> Since vect_recog_rotate_pattern has been extended to work also
> on signed types in r13-1100 we miscompile the testcase below.
> vect_recog_rotate_pattern actually emits correct scalar code into
> the pattern def sequence (in particular cast to utype, doing the
> 2 shifts in utype so that the right shift is logical and not arithmetic,
> or and then cast back to the signed type), but it didn't supply vectype
> for most of those pattern statements, which means that the generic handling
> fills it up later with the vectype provided by vect_recog_rotate_pattern.
> The problem is that it is vectype of the result of the whole pattern,
> i.e. vector of signed values in this case, while the conversion to utype,
> 2 shifts and or (everything with utype lhs in scalar code) should have
> uvectype as STMT_VINFO_VECTYPE.

What an interesting trap…

> Fixed with following patch, bootstrapped/regtested on x86_64-linux and
> i686-linux, ok for trunk?

Ok.

Thanks,
Richard 


> 2022-12-13  Jakub Jelinek  <jakub@redhat.com>
> 
>    PR tree-optimization/108064
>    * tree-vect-patterns.cc (vect_recog_rotate_pattern): Pass uvectype
>    as 4th argument to append_pattern_def_seq for statements with lhs
>    with utype type.
> 
>    * gcc.c-torture/execute/pr108064.c: New test.
> 
> --- gcc/tree-vect-patterns.cc.jj    2022-12-05 11:10:37.000000000 +0100
> +++ gcc/tree-vect-patterns.cc    2022-12-12 13:14:23.356628767 +0100
> @@ -3113,7 +3113,7 @@ vect_recog_rotate_pattern (vec_info *vin
>     {
>       def = vect_recog_temp_ssa_var (utype, NULL);
>       def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd0);
> -      append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
> +      append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
>       oprnd0 = def;
>     }
> 
> @@ -3137,7 +3137,7 @@ vect_recog_rotate_pattern (vec_info *vin
>     {
>       def = vect_recog_temp_ssa_var (utype, NULL);
>       def_stmt = gimple_build_assign (def, NOP_EXPR, oprnd1);
> -      append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
> +      append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
>     }
>   stype = TREE_TYPE (def);
> 
> @@ -3185,13 +3185,13 @@ vect_recog_rotate_pattern (vec_info *vin
>   def_stmt = gimple_build_assign (var1, rhs_code == LROTATE_EXPR
>                    ? LSHIFT_EXPR : RSHIFT_EXPR,
>                  oprnd0, def);
> -  append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
> +  append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
> 
>   var2 = vect_recog_temp_ssa_var (utype, NULL);
>   def_stmt = gimple_build_assign (var2, rhs_code == LROTATE_EXPR
>                    ? RSHIFT_EXPR : LSHIFT_EXPR,
>                  oprnd0, def2);
> -  append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt);
> +  append_pattern_def_seq (vinfo, stmt_vinfo, def_stmt, uvectype);
> 
>   /* Pattern detected.  */
>   vect_pattern_detected ("vect_recog_rotate_pattern", last_stmt);
> @@ -3202,7 +3202,7 @@ vect_recog_rotate_pattern (vec_info *vin
> 
>   if (!useless_type_conversion_p (type, utype))
>     {
> -      append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt);
> +      append_pattern_def_seq (vinfo, stmt_vinfo, pattern_stmt, uvectype);
>       tree result = vect_recog_temp_ssa_var (type, NULL);
>       pattern_stmt = gimple_build_assign (result, NOP_EXPR, var);
>     }
> --- gcc/testsuite/gcc.c-torture/execute/pr108064.c.jj    2022-12-12 13:22:29.875542508 +0100
> +++ gcc/testsuite/gcc.c-torture/execute/pr108064.c    2022-12-12 13:21:32.516377957 +0100
> @@ -0,0 +1,28 @@
> +/* PR tree-optimization/108064 */
> +
> +static inline short
> +foo (short value)
> +{
> +  return ((value >> 8) & 0xff) | ((value & 0xff) << 8);
> +}
> +
> +__attribute__((noipa))
> +void
> +bar (short *d, const short *s)
> +{
> +  for (unsigned long i = 0; i < 4; i++)
> +    d[i] = foo (s[i]);
> +}
> +
> +int
> +main ()
> +{
> +  short a[4] __attribute__((aligned (16))) = { 0xff, 0, 0, 0 };
> +  short b[4] __attribute__((aligned (16)));
> +  short c[4] __attribute__((aligned (16)));
> +
> +  bar (b, a);
> +  bar (c, b);
> +  if (a[0] != c[0])
> +    __builtin_abort ();
> +}
> 
>    Jakub
>